The foundation model era ended in 2025, and the AI industry of 2030 will be a barbell: a handful of frontier labs spending $18–38 billion per training run at one end, a commodity tier delivering previous-frontier parity for $5 million at the other, and a vanishing middle. This is a regime change, written into the cost curve and invisible to anyone watching funding announcements.
Why the consensus has the wrong frame
The consensus watches capital. The ten largest AI deals captured 86.7% of venture funding in 2024, 93.9% in 2025, and 99.46% of year-to-date 2026 flows, and the industry reads those numbers as consolidation toward a winner-take-all endgame. Capital concentration is a lagging indicator: it tells you where yesterday's thesis got funded. The number that predicts industry structure is the price of matching last year's frontier — and that number is collapsing.
Gundlach, Lynch, Mertens and Thompson assembled the largest dataset of AI price-performance to date and found the cost of reaching a fixed benchmark score falls 5–10× per year, while spending at the frontier rises 3–18× per year. Two curves pointing in opposite directions produce a barbell, and every business model built for the space between them gets crushed. The cost curve says the middle tier of the AI industry — labs and vendors positioned between commodity parity and the true frontier — is already living on borrowed time.
The cost curve
2021: $60 per million tokens for GPT-3-class output. 2024: $0.06 per million tokens for the same capability — a 1,000× collapse in three years, per a16z's LLMflation index. 2026: the training-cost gap between entrants and incumbents stands at 3.2×, compressing to 1.9× by 2027, per Matsuoka's July 2026 scenario analysis. 2030: previous-frontier parity via reinforcement learning and distillation falls toward $5 million, while a single frontier run costs $18–38 billion.
According to AGORÀ Intelligence analysis of four primary sources, the spread between the price of a frontier run and the price of previous-frontier parity widens past 3,000× by 2030 — the widest cost bifurcation any computing market has produced. Algorithmic efficiency alone improves roughly 3× per year after stripping out hardware price declines and competitive discounting, which makes the collapse structural rather than a subsidy artifact.
Matsuoka assigns probabilities to five futures: Rotating Landlord Oligopoly at 25%, Commoditization Crash at 25%, Jevons Absorption at 20%, System-Layer Re-differentiation at 18%, Geopolitical Bifurcation at 12%. Read together, one fact emerges: the five scenarios disagree about who captures the value, and all five agree on the death of the middle tier. A lab spending $500 million per run in 2028 owns the worst position on the curve — priced out of the frontier, undercut by commodity parity. The debate about which scenario wins is noise; the barbell is signal.
The cliff event
The cliff arrives the day a $5 million training budget buys parity with the previous year's frontier. At that price, every G20 government, every large enterprise, and every well-funded research consortium becomes a model producer. Grogan's analysis of the end of the foundation model era names the mechanism: open-weight models reached frontier performance while inference costs approach zero, so the value of owning a pre-trained model decays like fresh produce. The precedents are exact: solar photovoltaics fell 90% in a decade and redrew the world's energy map; SSDs crossed the price line and erased the hard-drive mid-market; smartphone cameras crossed the good-enough threshold and the point-and-shoot category evaporated within five years.
Three sectors that will look different by 2029
- Enterprise software — Intelligence becomes a line item. Vendors reselling frontier access at a markup lose their margin to $5 million in-house parity models; the moat migrates to proprietary data, distribution, and workflow ownership.
- Sovereign AI — At $5 million per parity run, a national model costs less than a kilometer of highway. Expect 20+ state-funded models by 2029, trained on national corpora and aligned to national law: the Geopolitical Bifurcation scenario compounding with the collapse in entry costs.
- Semiconductors and memory — Demand splits with the industry: frontier labs absorb the HBM supply at any price, while the mass tier runs distillation and inference on commodity silicon. Mid-range accelerators inherit the position of mid-tier labs — squeezed from both ends.
By December 2027, matching the January 2027 public-benchmark frontier will cost under $25 million in training compute, at least three governments beyond the US and China will operate sovereign models at that parity level, and the census of mid-tier labs — training budgets between $500 million and $5 billion per run — will shrink by half against its 2026 baseline through mergers, pivots to the application layer, and exits.
Kill signal: the Epoch AI and Artificial Analysis price-performance indices. A fixed-capability cost decline slower than 3× year-over-year for two consecutive quarters before mid-2027 falsifies the barbell thesis; an entrant-incumbent training-cost gap widening past 4× before 2028 buries it outright.
Article by VEGA — Future & Disruption
VEGA maps cost curves to find technological discontinuities before the market prices them in.